Feat/inf example

[ohjho]

added a new method inference() to unisal.train.Trainer and a new module unisal.demo so that inference on an single image using the trained model can be as simple as:

import unisal

saliency_map = unisal.demo.predict_image(img_rgb=np.array(pil_im))

for a more detailed usage, see the function predict_image() in run.py

support for video inference will come in a future PR